{"id":5627,"date":"2024-09-07T17:01:01","date_gmt":"2024-09-07T09:01:01","guid":{"rendered":""},"modified":"2024-09-07T17:01:01","modified_gmt":"2024-09-07T09:01:01","slug":"\u975e\u5e73\u8861\u91cd\u91c7\u6837\u7684\u76ee\u7684_csgo\u591a\u91cd\u91c7\u6837\u6297\u952f\u9f7f\u5f00\u4e0d\u5f00","status":"publish","type":"post","link":"https:\/\/mushiming.com\/5627.html","title":{"rendered":"\u975e\u5e73\u8861\u91cd\u91c7\u6837\u7684\u76ee\u7684_csgo\u591a\u91cd\u91c7\u6837\u6297\u952f\u9f7f\u5f00\u4e0d\u5f00"},"content":{"rendered":"

\u4e13\u680f\u7cfb\u5217\u6587\u7ae0 - \u77e5\u4e4e\u4e00\u3001\u804c\u4e1a\u89c4\u5212\u7bc7\u804c\u4e1a\u89c4\u5212\u4e0e\u9009\u62e9\u4e8c\u3001\u7b97\u6cd5\u9762\u7ecf\u7bc7\u67ab\u6866\uff1a\u6df1\u5ea6\u5b66\u4e60\u7b97\u6cd5\u5de5\u7a0b\u5e08\u9762\u7ecf(\u5fae\u8f6f\u3001\u963f\u91cc\u3001\u5546\u6c64\u3001\u6ef4\u6ef4\u3001\u534e\u4e3a\u3001\u6d77\u5eb7\u3001\u5e73\u5b89\u3001\u964c\u964c\u7b49offer)\u4e4b\u4e0a\u7bc7 \u67ab\u6866\uff1a\u6df1\u5ea6\u5b66\u4e60\u7b97\u6cd5\u5de5\u7a0b\u5e08\u9762\u7ecf(\u5fae\u8f6f\u3001\u963f\u91cc\u3001\u5546\u6c64\u3001\u6ef4\u6ef4\u3001\u534e\u4e3a\u3001\u6d77\u5eb7\u3001\u2026https:\/\/zhuanlan.zhihu.com\/p\/?<\/p>\n

\u76ee\u5f55<\/strong><\/p>\n

1. \u6982\u8ff0<\/p>\n

2. \u7c7b\u522b\u5e73\u8861\u91cd\u91c7\u6837<\/p>\n

3. Scheme-oriented sampling<\/p>\n

4.\u53c2\u8003\u8d44\u6599<\/p>\n

\n

1. \u6982\u8ff0<\/p>\n

\u5728\u300a\u4e0d\u5e73\u8861\u95ee\u9898: \u6df1\u5ea6\u795e\u7ecf\u7f51\u7edc\u8bad\u7ec3\u4e4b\u6b87\u300b\u4e00\u6587\u4e2d\uff0c\u7b14\u8005\u5df2\u5bf9\u7f13\u89e3\u4e0d\u5e73\u8861\u95ee\u9898\u7684\u65b9\u6cd5\u8fdb\u884c\u68b3\u7406\u3002\u9650\u4e8e\u7bc7\u5e45\u539f\u56e0\uff0c\u4ecb\u7ecd\u6bd4\u8f83\u7b3c\u7edf\u3002<\/p>\n

\u91cd\u91c7\u6837\u6cd5\u662f\u89e3\u51b3\u4e0d\u5e73\u8861\u95ee\u9898\u7684\u4e3b\u8981\u65b9\u6cd5\u4e4b\u4e00\uff0c\u5f88\u591a\u4eba\u7684\u7406\u89e3\u53ef\u80fd\u505c\u7559\u5728\u5bf9\u5934\u90e8\u7c7b\u522b\u8fdb\u884c\u6b20\u91c7\u6837\uff0c\u5bf9\u5c3e\u90e8\u7c7b\u522b\u8fdb\u884c\u8fc7\u91c7\u6837\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u91cd\u91c7\u6837\u5206\u4e3a\u968f\u673a\u8fc7\u91c7\u6837 (ROS, random over-sampling)\u548c\u968f\u673a\u6b20\u91c7\u6837 (RUS, random under-sampling)\uff0c\u5e7f\u6cdb\u5e94\u7528\u4e8e\u7c7b\u522b\u4e0d\u5e73\u8861\u95ee\u9898\u4e2d\u3002 ROS \u968f\u673a\u91cd\u590d\u5c3e\u90e8\u7c7b\u7684\u6837\u672c\uff0c\u800c RUS \u968f\u673a\u4e22\u5f03\u5934\u90e8\u7c7b\u7684\u6837\u672c\uff0c\u4ee5\u4f7f\u7c7b\u522b\u5e73\u8861\u3002 \u7136\u800c\uff0c\u5f53\u7c7b\u522b\u6781\u5ea6\u4e0d\u5e73\u8861\u65f6\uff0cROS \u503e\u5411\u4e8e\u8fc7\u5ea6\u62df\u5408\u5c3e\u90e8\u7c7b\uff0c\u800c RUS \u503e\u5411\u4e8e\u964d\u4f4e\u5934\u90e8\u7c7b\u7684\u6027\u80fd\u3002<\/p>\n

\u4e8b\u5b9e\u4e0a\uff0c\u9488\u5bf9\u91cd\u91c7\u6837\u65b9\u6cd5\u6709\u5f88\u591a\u7814\u7a76\uff0c\u5305\u62ec\u7c7b\u522b\u5e73\u8861\u91cd\u91c7\u6837\u548cScheme-oriented sampling\u3002<\/p>\n

2. \u7c7b\u522b\u5e73\u8861\u91cd\u91c7\u6837<\/h4>\n

2.1 Decoupling representation and classifier<\/p>\n

\u73b0\u6709\u7684\u4e0d\u5e73\u8861\u95ee\u9898\u89e3\u51b3\u65b9\u6848\u901a\u5e38\u91c7\u7528\u7c7b\u522b\u5e73\u8861\u7b56\u7565\uff0c\u4f8b\u5982\u901a\u8fc7\u635f\u5931\u91cd\u52a0\u6743\u3001\u6570\u636e\u91cd\u91c7\u6837\u6216\u4ece\u5934\u90e8\u7c7b\u5230\u5c3e\u90e8\u7c7b\u7684\u8fc1\u79fb\u5b66\u4e60\uff0c\u4f46\u5b83\u4eec\u4e2d\u7684\u5927\u591a\u6570\u90fd\u9075\u5faa\u8054\u5408\u5b66\u4e60\u7279\u5f81\u8868\u793a\u548c\u5206\u7c7b\u5668\u7684\u65b9\u6848\u3002\u5728\u8fd9\u9879\u5de5\u4f5c\u4e2d[2]\uff0c\u4f5c\u8005\u9996\u5148\u5bf9\u4e0d\u5e73\u8861\u8bc6\u522b\u4e2d\u7684\u5404\u79cd\u91c7\u6837\u7b56\u7565\u8fdb\u884c\u4e86\u5b9e\u8bc1\u7814\u7a76\uff0c\u91c7\u6837\u7b56\u7565\u5305\u62ec\u5b9e\u4f8b\u5e73\u8861\u91c7\u6837\u3001\u7c7b\u522b\u5e73\u8861\u91c7\u6837\u3001\u5e73\u65b9\u6839\u91c7\u6837\u548c\u6e10\u8fdb\u5e73\u8861\u91c7\u6837\uff0c\u5b9e\u4f8b\u5e73\u8861\u91c7\u6837\u662f\u6bcf\u4e2a\u6837\u672c\u88ab\u91c7\u6837\u7684\u6982\u7387\u76f8\u7b49\uff0c\u7c7b\u522b\u5e73\u8861\u91c7\u6837\u662f\u6bcf\u4e2a\u7c7b\u522b\u88ab\u91c7\u6837\u7684\u6982\u7387\u76f8\u7b49\uff1b\u5e73\u65b9\u6839\u91c7\u6837\u662f\u5b9e\u4f8b\u5e73\u8861\u91c7\u6837\u7684\u4e00\u79cd\u53d8\u4f53\uff0c\u5176\u4e2d\u6bcf\u4e2a\u7c7b\u522b\u7684\u91c7\u6837\u6982\u7387\u4e0e\u76f8\u5e94\u7c7b\u522b\u4e2d\u6837\u672c\u5927\u5c0f\u7684\u5e73\u65b9\u6839\u6709\u5173\uff1b \u6e10\u8fdb\u5e73\u8861\u91c7\u6837\u5728\u5b9e\u4f8b\u5e73\u8861\u91c7\u6837\u548c\u7c7b\u522b\u5e73\u8861\u91c7\u6837\u4e4b\u95f4\u8fdb\u884c\u6e10\u8fdb\u63d2\u503c\u3002\u7136\u540e\uff0c\u4f5c\u8005\u5c06\u5b66\u4e60\u8fc7\u7a0b\u89e3\u8026\u4e3a\u8868\u793a\u5b66\u4e60\u548c\u5206\u7c7b\u4e24\u9636\u6bb5\uff0c\u5e76\u7cfb\u7edf\u5730\u63a2\u7d22\u5728\u4e0d\u5e73\u8861\u95ee\u9898\u4e2d\uff0c\u4e0d\u540c\u7684\u5e73\u8861\u7b56\u7565\u5982\u4f55\u5f71\u54cd\u8fd9\u4e24\u4e2a\u9636\u6bb5\u3002\u7814\u7a76\u7ed3\u679c\u4ee4\u4eba\u60ca\u8bb6\uff1a\uff081\uff09\u6570\u636e\u4e0d\u5e73\u8861\u53ef\u80fd\u4e0d\u662f\u5b66\u4e60\u9ad8\u8d28\u91cf\u8868\u793a\u7684\u95ee\u9898\uff1b (2) \u4f7f\u7528\u6700\u7b80\u5355\u7684\u5b9e\u4f8b\u5e73\u8861\uff08\u81ea\u7136\uff09\u91c7\u6837\u5b66\u4e60\u5230\u7684\u7279\u79cd\u8868\u793a\uff0c\u4e5f\u53ef\u4ee5\u901a\u8fc7\u4ec5\u8c03\u6574\u5206\u7c7b\u5668\u6765\u5b9e\u73b0\u5f3a\u5927\u7684\u4e0d\u5e73\u8861\u8bc6\u522b\u80fd\u529b\u3002<\/p>\n

2.2 SimCal<\/p>\n

\"\u975e\u5e73\u8861\u91cd\u91c7\u6837\u7684\u76ee\u7684_csgo\u591a\u91cd\u91c7\u6837\u6297\u952f\u9f7f\u5f00\u4e0d\u5f00<\/p>\n

\u8bba\u6587[3]\u7cfb\u7edf\u5730\u7814\u7a76\u4e86\u6700\u5148\u8fdb\u7684\u4e24\u9636\u6bb5\u5b9e\u4f8b\u5206\u5272\u6a21\u578b Mask R-CNN \u5728\u6700\u8fd1\u7684\u957f\u5c3e LVIS \u6570\u636e\u96c6\u4e0a\u7684\u6027\u80fd\u4e0b\u964d\uff0c\u5e76\u63ed\u793a\u4e86\u4e00\u4e2a\u4e3b\u8981\u539f\u56e0\u662f\u6ca1\u6709\u5c06\u5bf9\u8c61\u63d0\u8bae (object proposals)\u51c6\u786e\u5206\u7c7b\u3002\u57fa\u4e8e\u8fd9\u6837\u7684\u89c2\u5bdf\uff0c\u4f5c\u8005\u9996\u5148\u8003\u8651\u5404\u79cd\u63d0\u9ad8\u4e0d\u5e73\u8861\u5206\u7c7b\u6027\u80fd\u7684\u6280\u672f\uff0c\u8fd9\u4e9b\u6280\u672f\u786e\u5b9e\u589e\u5f3a\u4e86\u5b9e\u4f8b\u5206\u5272\u7ed3\u679c\uff1b\u7136\u540e\u63d0\u51fa\u4e86\u4e00\u4e2a\u7b80\u5355\u7684\u6821\u51c6\u6846\u67b6 (Simple Calibration, SimCal\uff09\uff0c\u4e00\u79cd\u65b0\u7684\u53cc\u5c42\u7c7b\u5e73\u8861\u91c7\u6837\u7b56\u7565\u3002 \u5177\u4f53\u6765\u8bf4\uff0c\u53cc\u5c42\u91c7\u6837\u7b56\u7565\u7ed3\u5408\u4e86\u56fe\u50cf\u7ea7\u91cd\u91c7\u6837\u548c\u5b9e\u4f8b\u7ea7\u91cd\u91c7\u6837\uff0c\u4ee5\u7f13\u89e3\u5b9e\u4f8b\u5206\u5272\u4e2d\u7684\u7c7b\u522b\u4e0d\u5e73\u8861\u3002<\/p>\n

2.3 DCL<\/p>\n

\u52a8\u6001\u8bfe\u7a0b\u5b66\u4e60<\/strong>\uff08Dynamic curriculum learning, DCL\uff09[4]\u5f00\u53d1\u4e86\u4e00\u79cd\u65b0\u7684\u8bfe\u7a0b\u7b56\u7565\u6765\u52a8\u6001\u91c7\u6837\u6570\u636e\u4ee5\u8fdb\u884c\u7c7b\u522b\u518d\u5e73\u8861\u3002 \u5177\u4f53\u6765\u8bf4\uff0c\u968f\u7740\u8bad\u7ec3\u7684\u8fdb\u884c\uff0c\u4ece\u4e00\u4e2a\u7c7b\u4e2d\u91c7\u6837\u7684\u5b9e\u4f8b\u8d8a\u591a\uff0c\u8be5\u7c7b\u7684\u91c7\u6837\u6982\u7387\u5c31\u8d8a\u4f4e\u3002 \u6309\u7167\u8fd9\u4e2a\u601d\u8def\uff0cDCL\u9996\u5148\u8fdb\u884c\u968f\u673a\u62bd\u6837\u6765\u5b66\u4e60\u901a\u7528\u8868\u793a\uff0c\u7136\u540e\u6839\u636e\u8bfe\u7a0b\u7b56\u7565\u91c7\u6837\u66f4\u591a\u7684\u5c3e\u7c7b\u5b9e\u4f8b\u6765\u5904\u7406\u7c7b\u522b\u4e0d\u5e73\u8861\u3002<\/p>\n

2.4 Balanced meta-softmax<\/strong><\/p>\n

Balanced meta-softmax<\/strong> [5] \u5f00\u53d1\u4e86\u4e00\u79cd\u57fa\u4e8e\u5143\u5b66\u4e60\u7684\u91c7\u6837\u65b9\u6cd5\u6765\u4f30\u8ba1\u4e0d\u5e73\u8861\u5b66\u4e60\u4e2d\u4e0d\u540c\u7c7b\u522b\u7684\u6700\u4f73\u91c7\u6837\u7387\u3002 \u5177\u4f53\u6765\u8bf4\uff0c\u6240\u63d0\u51fa\u7684\u5143\u5b66\u4e60\u65b9\u6cd5\u662f\u4e00\u79cd\u53cc\u5c42\u4f18\u5316\u7b56\u7565\uff0c\u901a\u8fc7\u5728\u5e73\u8861\u7684\u5143\u9a8c\u8bc1\u96c6(meta validation set)\u4e0a\u4f18\u5316\u6a21\u578b\u5206\u7c7b\u6027\u80fd\u6765\u5b66\u4e60\u6700\u4f73\u6837\u672c\u5206\u5e03\u53c2\u6570\u3002<\/p>\n

2.5 FASA<\/p>\n

\u7279\u5f81\u589e\u5f3a\u548c\u91c7\u6837\u9002\u5e94<\/strong>\uff08Feature augmentation and sampling adaptation, FASA\uff09[6]\u63d0\u51fa\u4f7f\u7528\u5e73\u8861\u5143\u9a8c\u8bc1\u96c6\uff08\u4f5c\u4e3a\u5ea6\u91cf\uff09\u4e0a\u7684\u6a21\u578b\u5206\u7c7b\u635f\u5931\u6765\u8c03\u6574\u4e0d\u540c\u7c7b\u522b\u7684\u7279\u5f81\u91c7\u6837\u7387\uff0c\u4ece\u800c\u53ef\u4ee5\u5bf9\u4ee3\u8868\u6027\u4e0d\u8db3\u7684\u5c3e\u7c7b\u8fdb\u884c\u66f4\u591a\u91c7\u6837 .<\/p>\n

2.6 LOCE<\/p>\n

\u5177\u6709\u5206\u7c7b\u5e73\u8861\uff08LOCE\uff09\u7684\u4e0d\u5e73\u8861\u76ee\u6807\u68c0\u6d4b\u5668\uff08LOCE\uff09[7]\u63d0\u51fa\u4f7f\u7528\u5e73\u5747\u5206\u7c7b\u9884\u6d4b\u5206\u6570\uff08\u5373\u8fd0\u884c\u9884\u6d4b\u6982\u7387\uff09\u6765\u76d1\u63a7\u4e0d\u540c\u7c7b\u522b\u7684\u6a21\u578b\u8bad\u7ec3\uff0c\u5e76\u6307\u5bfc\u8bb0\u5fc6\u589e\u5f3a\u7279\u5f81\u91c7\u6837\u4ee5\u589e\u5f3a\u5c3e\u7c7b\u6027\u80fd\u3002<\/p>\n

2.7 VideoLT<\/p>\n

VideoLT [8] \u8bd5\u56fe\u89e3\u51b3\u4e0d\u5e73\u8861\u89c6\u9891\u8bc6\u522b\u95ee\u9898\uff0c\u5f15\u5165\u4e86\u4e00\u79cd\u65b0\u7684 FrameStack \u65b9\u6cd5\uff0c\u8be5\u65b9\u6cd5\u8fdb\u884c\u5e27\u7ea7\u91c7\u6837\u4ee5\u91cd\u65b0\u5e73\u8861\u7c7b\u5206\u5e03\u3002 \u5177\u4f53\u6765\u8bf4\uff0cFrameStack \u5728\u8bad\u7ec3\u65f6\u4f1a\u6839\u636e\u8fd0\u884c\u6a21\u578b\u7684\u6027\u80fd\u52a8\u6001\u8c03\u6574\u4e0d\u540c\u7c7b\u7684\u91c7\u6837\u7387\uff0c\u4f7f\u5176\u53ef\u4ee5\u4ece\u5c3e\u90e8\u7c7b\uff08\u901a\u5e38\u8fd0\u884c\u6027\u80fd\u8f83\u4f4e\uff09\u4e2d\u91c7\u6837\u66f4\u591a\u7684\u89c6\u9891\u5e27\uff0c\u4ece\u5934\u7c7b\u4e2d\u91c7\u6837\u66f4\u5c11\u7684\u5e27\u3002<\/p>\n

3. Scheme-oriented sampling<\/h4>\n

Scheme-oriented sampling\u8bd5\u56fe\u4e3a\u957f\u5c3e\u5b66\u4e60\u63d0\u4f9b\u4e00\u4e9b\u7279\u5b9a\u7684\u5b66\u4e60\u65b9\u6848\uff0c\u5982\u5ea6\u91cf\u5b66\u4e60\u548c\u96c6\u6210\u5b66\u4e60\u3002<\/p>\n

3.1 LMLE<\/p>\n

\"\u975e\u5e73\u8861\u91cd\u91c7\u6837\u7684\u76ee\u7684_csgo\u591a\u91cd\u91c7\u6837\u6297\u952f\u9f7f\u5f00\u4e0d\u5f00<\/p>\n

Large margin local embedding (LMLE)[9]\u4f7f\u7528\u4e86\u4e00\u79cd\u65b0\u7684\u4e94\u5143\u7ec4\u91c7\u6837\u65b9\u6848 (quintuplet sampling scheme)\uff0c\u4ee5\u5b66\u4e60\u4fdd\u6301inter-cluster\u548cinter-class margin\u7684\u9ad8\u8d28\u91cf\u7279\u5f81\u3002\u4e0d\u540c\u4e8e\u91c7\u7528\u4e24\u4e2a\u5bf9\u6bd4pair\u7684\u4e09\u5143\u7ec4\u635f\u5931 (triplet loss)\uff0cLMLE\u63d0\u51fa\u4e86\u4e00\u4e2a\u4e94\u5143\u7ec4\u91c7\u6837\u5668\u6765\u91c7\u6837\u56db\u4e2a\u5bf9\u6bd4pair\uff0c\u5305\u62ec\u4e00\u4e2a\u6b63\u6837\u672c\u5bf9\u548c\u4e09\u4e2a\u8d1f\u6837\u672c\u5bf9\uff0c\u5e76\u9f13\u52b1\u91c7\u6837\u7684\u4e94\u5143\u7ec4\u9075\u5faa\u7279\u5b9a\u7684\u8ddd\u79bb\u987a\u5e8f\u3002\u6b63\u6837\u672c\u5bf9\u7531\u951a\u70b9\u548c\u8ddd\u79bb\u951a\u70b9\u6700\u8fdc\u7684\u7c07\u5185\u6837\u672c\u7ec4\u6210\uff0c\u524d\u4e24\u4e2a\u8d1f\u6837\u672c\u5bf9\u6765\u81ea\u540c\u4e00\u7c7b\u522b\u5185\u8ddd\u79bb\u6700\u8fd1\u548c\u6700\u8fdc\u7684\u4e24\u4e2a\u7c07\u95f4\u6837\u672c\uff0c\u7b2c\u4e09\u4e2a\u8d1f\u6837\u672c\u5bf9\u6765\u81ea\u8ddd\u79bb\u6700\u8fd1\u7684\u7c7b\u95f4\u6837\u672c\u3002\u8fd9\u6837\uff0c\u5b66\u4e60\u5230\u7684\u8868\u793a\u4e0d\u4ec5\u7c7b\u5185\u95f4\u8ddd\u8f83\u5c0f\uff0c\u800c\u4e14\u7c7b\u95f4\u95f4\u8ddd\u8f83\u5927\u3002\u6b64\u5916\uff0c\u4e94\u5143\u7ec4\u635f\u5931\u4e2d\u7684\u6bcf\u4e2a\u6570\u636e\u6279\u6b21\u5305\u542b\u6765\u81ea\u4e0d\u540c\u7c7b\u522b\u7684\u76f8\u540c\u6570\u91cf\u7684\u6837\u672c\uff0c\u7528\u4e8e\u7c7b\u522b\u91cd\u5e73\u8861\u3002<\/p>\n

(\u7efc\u8ff0\u4e2dLMLE\u63cf\u8ff0\u6709\u95ee\u9898\uff0c\u6839\u636e\u539f\u6587\u8fdb\u884c\u4fee\u6b63\uff1b\u8fd9\u91cc\u6211\u6709\u70b9\u597d\u5947\uff0c\u4e3a\u4ec0\u4e48\u4e0d\u79f0\u4e3a\u4e09\u4e2a\u6b63\u6837\u672c\u5bf9\uff0c\u4e00\u4e2a\u8d1f\u6837\u672c\u5bf9)<\/p>\n

3.2 PRS<\/p>\n

Partitioning reservoir sampling (PRS) [10]\u63d0\u51fa\u4e86\u4e00\u79cd\u57fa\u4e8e\u91cd\u653e\u7684\u91c7\u6837\u65b9\u6cd5\u6765\u5904\u7406\u6301\u7eed\u7684\u957f\u5c3e\u5b66\u4e60\u3002 \u4e00\u4e2a\u5173\u952e\u6311\u6218\u662f\u56de\u653e\u8bb0\u5fc6\u65e0\u6cd5\u8003\u8651\u7c7b\u522b\u4e0d\u5e73\u8861\u7684\u95ee\u9898\uff0c\u56e0\u4e3a\u6ca1\u6709\u5173\u4e8e\u672a\u6765\u8f93\u5165\u7684\u4fe1\u606f\u53ef\u7528\u3002 \u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff0cPRS \u5f00\u53d1\u4e86\u4e00\u79cd\u5728\u7ebf\u5185\u5b58\u7ef4\u62a4\u7b97\u6cd5\uff0c\u53ef\u4ee5\u52a8\u6001\u7ef4\u62a4\u6765\u81ea\u4e0d\u540c\u7c7b\u522b\u7684\u6837\u672c\u7684\u8fd0\u884c\u7edf\u8ba1\u4fe1\u606f\u3002 PRS \u53ef\u4ee5\u6839\u636e\u8fd0\u884c\u7edf\u8ba1\u52a8\u6001\u8c03\u6574\u4e0d\u540c\u7c7b\u7684\u5185\u5b58\u5927\u5c0f\u548c\u91c7\u6837\u8f93\u5165\/\u8f93\u51fa\u64cd\u4f5c\u7684\u65b9\u6848\u3002<\/p>\n

3.3 BBN<\/p>\n

\"\u975e\u5e73\u8861\u91cd\u91c7\u6837\u7684\u76ee\u7684_csgo\u591a\u91cd\u91c7\u6837\u6297\u952f\u9f7f\u5f00\u4e0d\u5f00<\/p>\n

\u53cc\u8fb9\u5206\u652f\u7f51\u7edc\uff08BBN\uff09[11] \u5f00\u53d1\u4e86\u4e24\u4e2a\u7f51\u7edc\u5206\u652f\uff08\u5373\u4f20\u7edf\u5b66\u4e60\u5206\u652f\u548c\u91cd\u65b0\u5e73\u8861\u5206\u652f\uff09\uff0c\u4ee5\u57fa\u4e8e\u65b0\u7684\u53cc\u8fb9\u91c7\u6837\u7b56\u7565\u5904\u7406\u7c7b\u4e0d\u5e73\u8861\u3002 \u5177\u4f53\u6765\u8bf4\uff0cBBN \u5bf9\u5e38\u89c4\u5206\u652f\u5e94\u7528\u5747\u5300\u62bd\u6837\u6765\u6a21\u62df\u539f\u59cb\u7684\u957f\u5c3e\u8bad\u7ec3\u5206\u5e03\uff1b \u540c\u65f6\uff0cBBN \u5bf9\u518d\u5e73\u8861\u5206\u652f\u5e94\u7528\u4e86\u4e00\u4e2a\u53cd\u5411\u91c7\u6837\u5668\uff0c\u4ee5\u91c7\u6837\u66f4\u591a\u7684\u5c3e\u7c7b\u6837\u672c\uff0c\u4ee5\u63d0\u9ad8\u5c3e\u7c7b\u6027\u80fd\u3002 \u6700\u7ec8\u7684\u9884\u6d4b\u662f\u4e24\u4e2a\u7f51\u7edc\u5206\u652f\u7684\u52a0\u6743\u548c\u3002 \u4e4b\u540e\uff0c\u957f\u5c3e\u591a\u6807\u7b7e\u89c6\u89c9\u8bc6\u522b\uff08LTML\uff09[12]\u6269\u5c55\u4e86\u53cc\u8fb9\u5206\u652f\u7f51\u7edc\u4ee5\u89e3\u51b3\u957f\u5c3e\u591a\u6807\u7b7e\u5206\u7c7b\u95ee\u9898\u3002 \u51e0\u4f55\u7ed3\u6784\u8f6c\u79fb\uff08GIST\uff09[13] \u8fd8\u63a2\u7d22\u4e86\u8fd9\u79cd\u53cc\u8fb9\u91c7\u6837\u7b56\u7565\uff0c\u7528\u4e8e\u4ece\u5934\u5230\u5c3e\u7684\u77e5\u8bc6\u8f6c\u79fb\u3002<\/p>\n

3.4 BAGS<\/p>\n

balanced group softmax (BAGS) [14] \u63d0\u51fa\u6839\u636e\u6bcf\u4e2a\u7c7b\u4e2d\u7684\u6837\u672c\u6570\u91cf\u5c06\u7c7b\u5212\u5206\u4e3a\u51e0\u4e2a\u5e73\u8861\u7ec4\uff0c\u5176\u4e2d\u6bcf\u4e2a\u7ec4\u5177\u6709\u76f8\u4f3c\u6570\u91cf\u7684\u8bad\u7ec3\u6570\u636e\u7684\u7c7b\u3002 \u5728\u6b64\u4e4b\u540e\uff0cBAGS \u4f7f\u7528\u4e0d\u540c\u7684\u6837\u672c\u7ec4\u6765\u8bad\u7ec3\u4e0d\u540c\u7684\u5206\u7c7b\u5934\uff0c\u4ee5\u4fbf\u5b83\u4eec\u5bf9\u5177\u6709\u76f8\u4f3c\u6570\u91cf\u7684\u8bad\u7ec3\u6570\u636e\u7684\u7c7b\u6267\u884c softmax \u64cd\u4f5c\uff0c\u4ece\u800c\u907f\u514d\u7531\u4e8e\u4e0d\u5e73\u8861\u800c\u5bfc\u81f4\u4e25\u91cd\u504f\u5dee\u7684\u5206\u7c7b\u5668\u3002<\/p>\n

3.5 LST<\/p>\n

learning to segment the tail (LST) [15]\u8fd8\u5c06\u8bad\u7ec3\u6837\u672c\u5206\u6210\u51e0\u4e2a\u5e73\u8861\u7684\u5b50\u96c6\uff0c\u5e76\u57fa\u4e8e\u7c7b\u589e\u91cf\u5b66\u4e60\u5904\u7406\u6bcf\u4e2a\u5b50\u96c6\u3002 \u4e3a\u4e86\u89e3\u51b3\u7c7b\u589e\u91cf\u5b66\u4e60\u8fc7\u7a0b\u4e2d\u7684\u707e\u96be\u6027\u9057\u5fd8\uff0cLST \u5f00\u53d1\u4e86\u4e00\u79cd\u7c7b\u5e73\u8861\u7684\u6570\u636e\u56de\u590d\/\u91c7\u6837\u7b56\u7565\uff0c\u8be5\u7b56\u7565\u4e3a\u77e5\u8bc6\u84b8\u998f\u4fdd\u6301\u76f8\u5bf9\u5e73\u8861\u7684\u6837\u672c\u96c6\u3002<\/p>\n

3.6 ACE<\/p>\n

ally complementary experts (ACE) [16] \u4e0d\u662f\u5c06\u6837\u672c\u5212\u5206\u4e3a\u51e0\u4e2a\u5e73\u8861\u7684\u7ec4\uff0c\u800c\u662f\u5c06\u6837\u672c\u5212\u5206\u4e3a\u51e0\u4e2a\u6280\u80fd\u591a\u6837\u5316\u7684\u5b50\u96c6\uff0c\u5176\u4e2d\u4e00\u4e2a\u5b50\u96c6\u5305\u542b\u6240\u6709\u7c7b\uff0c\u4e00\u4e2a\u5305\u542b\u4e2d\u95f4\u7c7b\u548c\u5c3e\u90e8\u7c7b\uff0c\u53e6\u4e00\u4e2a\u53ea\u5305\u542b\u5c3e\u90e8\u7c7b\u3002 \u57fa\u4e8e\u8fd9\u4e9b\u5b50\u96c6\uff0cACE \u57f9\u8bad\u4e0d\u540c\u7684\u4e13\u5bb6\uff0c\u4f7f\u5176\u5177\u5907\u7279\u5b9a\u548c\u4e92\u8865\u7684\u96c6\u6210\u5b66\u4e60\u6280\u80fd\u3002<\/p>\n

4.\u53c2\u8003\u8d44\u6599<\/h4>\n

[1] \u4e0d\u5e73\u8861\u95ee\u9898: \u6df1\u5ea6\u795e\u7ecf\u7f51\u7edc\u8bad\u7ec3\u4e4b\u6b87<\/p>\n

[2] B. Kang, S. Xie, M. Rohrbach, Z. Yan, A. Gordo, J. Feng, and Y. Kalantidis, \u201cDecoupling representation and classifier for long-tailed recognition,\u201d in International Conference on Learning Representations, 2020.<\/p>\n

[3] T. Wang, Y. Li, B. Kang, J. Li, J. Liew, S. Tang, S. Hoi, and J. Feng, \u201cThe devil is in classification: A simple framework for long-tail instance segmentation,\u201d in European Conference on Computer Vision, 2020.<\/p>\n

[4] Y. Wang, W. Gan, J. Yang, W. Wu, and J. Yan, \u201cDynamic curriculum learning for imbalanced data classification,\u201d in International Conference on Computer Vision, 2019, pp. 5017\u20135026.<\/p>\n

[5] R. Jiawei, C. Yu, X. Ma, H. Zhao, S. Yi et al., \u201cBalanced meta-softmax for long-tailed visual recognition,\u201d in Advances in Neural Information Processing Systems, 2020.<\/p>\n

[6] Y. Zang, C. Huang, and C. C. Loy, \u201cFasa: Feature augmentation and sampling adaptation for long-tailed instance segmentation,\u201d in International Conference on Computer Vision, 2021.<\/p>\n

[7] C. Feng, Y. Zhong, and W. Huang, \u201cExploring classification equilibrium in long-tailed object detection,\u201d in International Conference on Computer Vision, 2021.<\/p>\n

[8] X. Zhang, Z. Wu, Z. Weng, H. Fu, J. Chen, Y.-G. Jiang, and L. Davis, \u201cVideolt: Large-scale long-tailed video recognition,\u201d in International Conference on Computer Vision, 2021.<\/p>\n

[9] C. Huang, Y. Li, C. C. Loy, and X. Tang, \u201cLearning deep representation for imbalanced classification,\u201d in Computer Vision and Pattern Recognition, 2016.<\/p>\n

[10] C. D. Kim, J. Jeong, and G. Kim, \u201cImbalanced continual learning with partitioning reservoir sampling,\u201d in European Conference on Computer Vision, 2020, pp. 411\u2013428.<\/p>\n

[11] B. Zhou, Q. Cui, X.-S. Wei, and Z.-M. Chen, \u201cBbn: Bilateral-branch network with cumulative learning for long-tailed visual recognition,\u201d in Computer Vision and Pattern Recognition, 2020, pp. 9719\u20139728.<\/p>\n

[12] H. Guo and S. Wang, \u201cLong-tailed multi-label visual recognition by collaborative training on uniform and re-balanced samplings,\u201d in Computer Vision and Pattern Recognition, 2021, pp. 15 089\u201315 098.<\/p>\n

[13] B. Liu, H. Li, H. Kang, G. Hua, and N. Vasconcelos, \u201cGistnet: a geometric structure transfer network for long-tailed recognition,\u201d in International Conference on Computer Vision, 2021.<\/p>\n

[14] Y. Li, T. Wang, B. Kang, S. Tang, C. Wang, J. Li, and J. Feng, \u201cOvercoming classifier imbalance for long-tail object detection with balanced group softmax,\u201d in Computer Vision and Pattern Recognition, 2020, pp. 10 991\u201311 000.<\/p>\n

[15] X. Hu, Y. Jiang, K. Tang, J. Chen, C. Miao, and H. Zhang, \u201cLearning to segment the tail,\u201d in Computer Vision and Pattern Recognition, 2020.<\/p>\n

[16] J. Cai, Y. Wang, and J.-N. Hwang, \u201cAce: Ally complementary experts for solving long-tailed recognition in one-shot,\u201d in International Conference on Computer Vision, 2021.<\/p>\n","protected":false},"excerpt":{"rendered":"\u975e\u5e73\u8861\u91cd\u91c7\u6837\u7684\u76ee\u7684_csgo\u591a\u91cd\u91c7\u6837\u6297\u952f\u9f7f\u5f00\u4e0d\u5f00\u4e8b\u5b9e\u4e0a\uff0c\u9488\u5bf9\u91cd\u91c7\u6837\u65b9\u6cd5\u6709\u5f88\u591a\u7814\u7a76\uff0c\u5305\u62ec\u7c7b\u522b\u5e73\u8861\u91cd\u91c7\u6837\u548cScheme-orientedsampling","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[],"tags":[],"_links":{"self":[{"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/posts\/5627"}],"collection":[{"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/comments?post=5627"}],"version-history":[{"count":0,"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/posts\/5627\/revisions"}],"wp:attachment":[{"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/media?parent=5627"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/categories?post=5627"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/tags?post=5627"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}